Amazon's tokenmaxxing memo: token usage is a flawed AI metric

Amazon has told its software engineers to stop treating AI usage as proof of productivity. In a memo now public on its own site, the company calls token usage and AI-tool adoption "deeply flawed" metrics and says tokenmaxxing "is at odds with frugality." When one of the biggest AI vendors on earth tells its own staff to burn fewer tokens, small teams paying retail for the same tokens should listen.

What actually happened

CNBC reported on October 27 that the memo went around internally in September and was posted publicly on Tuesday. It is signed by Terk Terkowitz, a VP and Distinguished Engineer for Worldwide Amazon Stores, and it rewrites how Amazon defines the role of its software engineers and managers. The key points:

• Input metrics are flawed. Token usage, adoption of AI tools, and throughput counts like code changes and reviews can drive bad outcomes when they become targets. They "correlate well" with useful output for now, so the memo says use them "with care and judgement."
• Watch both tails. Leaders should look at the 10th percentile of AI adoption to learn why some teams are not using it, and at the 90th to check that heavy use comes with more delivery.
• "A finding, not a failing." If AI does not help a piece of work, engineers should write down why and share it.
• Outcomes over tools. The memo says neither the developer role nor its output "is defined by the use of AI."

Context from CNBC: Amazon shut down an internal token leaderboard in May after staff gamed it, and a spokesperson said Amazon has no caps on employee token usage.

Why token usage is the wrong AI metric for your business

We see the small-business version of this all the time. A team signs up for three AI coding tools, the bill doubles, and the only proof of value is a usage dashboard. Usage is a cost line. It is not a result.

How we would measure instead:

1. Pick one outcome per workflow. Tickets closed, quotes sent, bugs shipped to production. Not tokens.
2. Put cost next to it. Monthly AI spend divided by that outcome. If the ratio climbs, something is wrong.
3. Review the top spender. Your 90th percentile user should also be your fastest shipper. If not, find out why.
4. Log the misses. When AI did not help, write it down. That note saves the next person a week.

Sources: Amazon, CNBC.